Remote Sens. 2011, 3, 1323-1343; doi:10.3390/rs3071323OPEN ACCESS

Remote SensingISSN 2072-4292

www.mdpi.com/journal/remotesensing

Article

Accuracy Enhancement of ASTER Global Digital ElevationModels Using ICESat DataHossein Arefi ? and Peter Reinartz

Remote Sensing Technology Institute, German Aerospace Center (DLR), D-82234 Wessling, Germany;E-Mail: peter.reinartz@dlr.de

? Author to whom correspondence should be addressed; E-Mail: hossein.arefi@dlr.de;Tel.: +49-8153-282165; Fax: +49-8153-281444.

Received: 28 April 2011; in revised form: 27 June 2011 / Accepted: 27 June 2011 /Published: 1 July 2011

Abstract: Global Digital Elevation Models (GDEM) are considered very attractive forcurrent research and application areas due to their free and wide range accessibility. TheASTER Global Digital Elevation Model exhibits the highest spatial resolution data of allglobal DEMs and it is generated for almost the whole globe. Unfortunately, ASTERGDEM data include many artifacts and height errors that decrease the quality and elevationaccuracy significantly. This study provides a method for quality improvement of theASTER GDEM data by correcting systematic height errors using ICESat laser altimetrydata and removing artifacts and anomalies based on a segment-based outlier detection andelimination algorithm. Additionally, elevation errors within water bodies are corrected usinga water mask produced from a high-resolution shoreline data set. Results indicate that theaccuracy of the corrected ASTER GDEM is significantly improved and most artifacts areappropriately eliminated. Nevertheless, artifacts containing lower height values with respectto the neighboring ground pixels are not entirely eliminated due to confusion with some realnon-terrain 3D objects. The proposed method is particularly useful for areas where otherhigh quality DEMs such as SRTM are not available.

Keywords: Digital Elevation Models (DEM); ASTER; geodesic image reconstruction;outlier removal; water mask; enhancement

Remote Sens. 2011, 3 1324

1. Introduction

Global Digital Elevation Models (GDEM) which are freely available for all users and accessible viathe Internet are used for many applications in several disciplines. GDEMs are provided in differentresolutions, are generated by employing various techniques, and exhibit different ranges of coverage onthe earth’s surface. Recently, in July 2009, the Ministry of Economy, Trade, and Industry (METI) ofJapan and the United States National Aeronautics and Space Administration (NASA) released a newGDEM, which is a great step towards a worldwide high-resolution elevation model. It is produced fromAdvanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) stereo images. The aimof the ASTER GDEM is to provide a highly accurate DEM, which covers the complete landmass, andis freely available to all users. Furthermore it is aimed to be used as a platform for analysis of data inthe fields of disaster monitoring (e.g., volcanic or flood hazard map), hydrology (e.g., water resourcemanagement), energy (e.g., oil resource exploration), and stereoscopic visualization (e.g., for Bird’s-eyeviews and flight simulations) [1]. The ASTER GDEM is generated with a spatial resolution of about 30 m(1 arc-second) from the original 15 m ASTER image ground sampling distance (GSD) in the horizontalplane. Therefore, the ASTER GDEM is the highest resolution DEM among the free accessible globalDEMs. Other important GDEMs include the Shuttle Radar Topography Mission (SRTM) DEM with aresolution of 3 arc-seconds (∼90 m at the equator) and the Global Topographic Data (GTOPO30) with aresolution of 30 arc-seconds (∼1,000 m at the equator).

The ASTER GDEM covers land surfaces between −83◦ and +83◦. Moreover, the ASTER GDEMdata are available also for those high latitude and steep mountainous areas that are not covered bySRTM (available between−58◦ and +60◦) due to orbit restrictions, radar shadowing, and foreshorteningeffects [1].

The vertical and horizontal accuracies of ASTER GDEM are estimated in pre-production level at95% confidence as 20 m, and 30 m respectively. Prior to releasing the ASTER GDEM data to theglobal user community in July 2009, an extensive preliminary validation study in cooperation with theUS Geological Survey (USGS), ERDAS, and other investigators has been performed [2]. The results ofthe accuracy assessment by this study prove that the pre-production estimated vertical accuracy of 20 mat 95% confidence is globally correct. Looking to single tiles of size 1◦ × 1◦, some exhibit a verticalaccuracy of better than 20 m but for others the accuracy can be worse. The study concludes that themajor errors of the GDEM are the following issues:

• The ASTER GDEM shows almost for all areas a general negative bias of about 5 m on average.The comparison in the Continental United States (CONUS) proves that the ASTER GDEM is onaverage about 4 m lower than the National Elevation Dataset (NED), and more than 6 m lower inopen and flat regions.

• The GDEMs data contain artifacts and anomalies, which can produce large elevation errors onlocal scales. These errors occur mainly in areas where just few stereo images are available orin the regions with persistent “clouds”. Artifacts due to existing small numbers of stereo imagesappear as straight lines, “pits”, “bumps”, “mole-runs”, and other geometric shapes [2].

Remote Sens. 2011, 3 1325

• In the ASTER GDEM, no inland water mask has been applied and therefore the elevation valueson most of the lakes are inaccurate, i.e., the elevation values of water body regions are not set to asingle “flattened” height value.

Therefore, the major motivation of the work in this paper is to provide methodologies to generatean improved and thus more appropriate digital elevation model, in particular for the areas wherethe SRTM data are not available, i.e., above 60◦ North and below 58◦ South. Accordingly, in thispaper the quality of the ASTER GDEM is improved. This is performed by using the ICESat laseraltimetry data for absolute height improvement and some local refinement measures. Huber et al. [3]proposed an algorithm for correcting the height accuracy of the SRTM DEMs using the ICESat laseraltimetry points. The Geoscience Laser Altimeter System (GLAS) instrument mounted on ICESat(Ice, Cloud, and land Elevation Satellite), measures elevation data since 2003. The absolute heightaccuracy of the laser points varies depending on land cover and relief and it is generally better than1 m [4]. GLAS produces a series of approximately 70 m diameter footprints that are separated bynearly 170 m intervals along track and 30 km across track (at the equator) and 5 km at 80◦ latitude [5].Huber et al. [3] proposed a strategy to refine and extract high quality laser points from the ICESatwaveform dataset to be applied for SRTM DEM correction. They investigate the deviation of ICEsatpoints from a reference Digital Surface Model (DSM) (with accuracy of ±0.5 m in open areas and±1.5 m in areas covered with vegetation) and a reference Digital Terrain Model (DTM) (with accuracyof ±0.5 m). As final refinement, the selected parameters of the laser waveform, i.e., received energy,signal width, peak numbers, and standard deviation, has been adapted to attain highly accurate laserpoints. The final height accuracy of the selected ICESat points after employing the predefined criteriaare obtained as 0.64 m mean and 1.30 m standard deviation regarding the investigated 65 points. Asfurther work, they have proposed an algorithm to enhance the global elevation quality of SRTM datausing the selected ICESat points. It is assumed that the SRTM data contains long wavelength errors upto a level of 10 m [6]. Therefore, a model for long wavelength errors based on spherical harmonics hasbeen generated in which the coefficients of the model are estimated by a least squares adjustment basedon the differences between ISESat and SRTM. The method based on spherical harmonics correctionfunction calculates continuously the error offset to SRTM rather than adding only one offset value foreach tile.

While spherical harmonic functions are suitable in case of solving long wavelength errors, this is notthe case for the ASTER GDEM. The ASTER GDEM is based on a group of individual acquisitions andnot on continuously acquired stripes as it is the case for SRTM DEM.

The overall goal of this paper is to enhance the quality of the ASTER GDEM with focus on theregions where SRTM DEM is not available. Hence, the paper has the following objectives:

• Correction of the height offset errors by employing selected ICESat laser altimetry data as controlpoints. A correction layer is generated by measuring the deviation of the ASTER GDEM from thecorresponding ICESat laser point.

• Removal of local artifacts and anomalies using a segmentation-based method.

• Modifying the height of pixels of the lakes and water bodies by employing a water mask layerprovided by an existing freely accessible water boundary database.

Remote Sens. 2011, 3 1326

2. Methods

According to the errors reported in the first ASTER GDEM validation report [2], summarized inSection 1, a methodology is proposed in this paper to reduce and correct the three errors in the followingthree steps:

2.1. Correction of the Height Errors Using ICESat Laser Altimetry Points:

The focus is on the refinement of the elevation data by reducing the ASTER GDEM bias based ona correction layer, which is provided from the ICESat laser altimetry data. For that, a correction heightlayer is provided according to the height deviation of the ASTER GDEM points from the correspondingICESat points.

As first step, the ICESat point clouds corresponding to the test area are extracted from the datasetrepository. In order to use the ICESat points as Ground Control Points (GCPs) to be applied onevaluation and correction of the DEMs, the erroneous points caused by clouds, outliers, underlyingslopes or vegetation should be eliminated first. Based on the criteria proposed by Huber et al. [3], thepoints containing errors are filtered to reach a vertical accuracy of nearly 1 m compared to the referencedata. Accordingly, the ICESat waveform laser points are filtered by selecting appropriate thresholds fordifferent parameters as follows: (1) number of peaks as a criterion to distinguish bare soil and forestareas, (2) received energy from signal begin to signal end, and (3) the signal width (in meter), which isthe distance between signal begin and signal end.

Figure 1 represents the locations of the ICESat laser altimetry data over a sample of ASTER GDEMdata. It additionally shows the direction and magnitude of the height deviation of the GDEM regardingthe ICESat points. The direction of the vertical arrows proves that the ASTER GDEM is locatedgenerally below the correct elevation. The original laser points located in the region are refined using theparameters recommended by Huber et al. [3] as follows: Points with less than 6 peaks, received energylower than 10 fJ (femtojoule), and a signal width smaller than 25 m are selected. The ICESat points thatare not satisfying all the predefined criteria are eliminated from the dataset. They are visualized as bluearrows in Figure 1.

After selecting the corresponding ICESat points and their correction, in order to obtain high accuracypoints, analogous ASTER GDEM points are extracted using interpolation. An additional criterion isutilized to reject the ICESat laser points that are not consistent with the GDEM points considering themas outliers. In this step, the aim is to obtain the vertical offsets related to each ASTER GDEM pointto reduce the negative bias of the ASTER GDEM. For that purpose the difference between the selectedICESat points and their corresponding ASTER GDEM points are measured. A correction height layer isprovided by interpolation of the height differences within the test area (cf. Figure 2).

In practice, in order to increase the quality of the correction layer and to be able to apply the algorithmmore globally, an area much larger than the test area is analyzed for extracting ICESat points as well asthe corresponding GDEM points. This procedure has two benefits: (1) since the ICESat points are not asdense as the GDEM points, selecting a larger area allows to more precisely measure the height offset forthe corresponding area, and (2) extracting the ICESat points for a larger area and applying the correctionlayer only on the test region provides a smooth transition on overlapping areas between neighboring

Remote Sens. 2011, 3 1327

scenes. For the interpolation of the height differences over the entire scene, an ordinary Moving Averagemethod is employed. It assigns the values to the grid nodes by averaging the data within a search ellipsearound each node.

Figure 1. ICESat points located on the ASTER GDEM, corresponding to an area of6 km × 5.1 km plus vector arrows showing the vertical errors of each point. Refined ICESatpoints to get points located only on the ground surface (red arrows) as well as eliminatedpoints (blue arrows) are visualized.

2620 2640 2660 2680 2700 2720 2740 2760

1800

1850

1900

1950210

240

270

300

330

360

390

420

2.2. Segment-Based Artifacts and Anomalies Detection and Elimination:

In this step, an algorithm is proposed to detect and remove artifacts and anomalies as outliers fromthe ASTER GDEM based on a segmentation. The algorithm to extract and eliminate the artifacts andanomalies from the ASTER GDEM using segmentation technique has been reported in an earlier workof the authors [7].

Remote Sens. 2011, 3 1328

Figure 2. Elevation correction layer corresponding to an area of about 30 km × 30 km. Thevalues on the X and Y axis represent the pixel coordinates.

-15

-14.5

-14.

5

-14

-14

-14

-14

-14

-13.5

-13.

5-1

3.5

-13.

5

-13

-13

-13

-13-1

2.5

-12.5

-12.5

-12.5

-12

-12

-12-12

-12

100 200 300 400 500 600 700 800 900 1000

100

200

300

400

500

600

700

800

900

1000

In order to extract the outliers effectively, their types and specifications are identified first. Thespecifications of the common errors existing in the released ASTER GDEM, which are detailed inASTER GDEM validation report [2], are summarized as follows:

• “Pits” occur as small negative elevation anomalies, which vary from a few meters to about 100 min height. Figure 3 illustrates an example of the area containing the “pit” artifacts and theirassociation with stack number boundaries (cf. Figure 3(B)). Stack numbers correspond to thenumber of image pairs that are using to generate DEM.

Figure 3. Example of “pit” artifacts; (A) shaded relief, (B) clear relation to the stack (scene)number boundaries, (C) appearance in ASTER GDEM represented as gray scale image(originally appeared in [8]).

Remote Sens. 2011, 3 1329

• “Bumps” appear as positive elevation anomaly artifact; their magnitude can range from just fewmeters to more than 100 m in height (cf. Figure 4).

Figure 4. Example of “bump” artifacts; (A) shaded relief, (B) clear relation to the stacknumber boundaries, (C) appearance in ASTER GDEM image (originally appeared in [8]).

• “Mole runs” are curvilinear anomalies above the ground, which are less common than pits andbumps and occur in relatively flat terrains. The corresponding magnitude of “mole runs” is muchless than for the two previous anomalies and it ranges from barely perceptible to a few meters, andrarely more than 10 m (cf. Figure 5). Due to their linear behavior, they can be easily recognized ina shaded DEM.

Figure 5. Example of “mole-run” artifacts; (A) shaded relief, (B) clear relation to the stacknumber boundaries, (C) less obvious in ASTER GDEM image (originally appeared in [8]).

The artifacts and anomalies produced by small and different stack numbers are apparent in almost allASTER GDEM tiles. In addition, the effects of residual clouds have been already eliminated in Version1 of the ASTER GDEM by replacing the elevation with −9999 values [2]. Here, an algorithm based onimage reconstruction using geodesic morphological dilation [9,10] is employed to extract the regionalextrema, which is later used for eliminating the “pits” and “bumps”. Geodesic dilation from gray-scalemathematical morphology differs to basic dilation where an image and a structuring element are involvedin the filtering process. In geodesic dilation the dilated image is additionally “masked” with a predefined

Remote Sens. 2011, 3 1330

“mask” image. Equation (1) shows the geodesic dilation of image J (marker) using mask I. In mostapplications, the marker image is defined by a height offset to the mask image, which generally representsthe original DEM. Figure 6 illustrates the difference between geodesic and basic image dilation as wellas reconstruction based on geodesic dilation in a profile view of a simple building with gable roof. Theinput image (a), here called marker, is enlarged by dilation, i.e., the gray region in (b), and limited by themask image (I). The result of geodesic dilation is shown in (d) with a dashed line around it depicting themask image. If this process, i.e., dilation and limitation by mask, is iteratively continued, it stops afterfour iterations reaching stability. The result provided by this step is called reconstruction of marker (J)by mask (I) using geodesic dilation (cf. Figure 6(g)). The number of iterations, i.e., n in Equation (2), tocreate reconstructed image varies from one sample to another. In the example presented in Figure 6 thereconstruction procedure stops after four iterations.

Figure 6. Geodesic dilation.

Accordingly, geodesic dilation (δI ) and image reconstruction are defined as

δ(1)I (J) = (J

⊕B)

∧I, (1)

δ(n)I (J) = δ

(1)I (J) ◦ δ(1)

I (J) ◦ .... ◦ δ(1)I (J)︸ ︷︷ ︸

n times

(2)

Equation (2) defines the morphological reconstruction of the marker image (J) based on geodesicdilation (δI) (cf. Equation (1)) if the iterative geodesic dilations reaches to stability. The basic dilation(δ) of marker and point-wise minimum (∧) between dilated image and mask (I) is employed iterativelyuntil stability. Looking at the reconstructed image of the example depicted in Figure 6 shows thatthe upper part of the object, i.e., the difference between marker and mask is suppressed during imagereconstruction. Therefore, the result of gray scale reconstruction depends on the height offset betweenthe marker and the mask images and accordingly, different height offset suppress different parts of theobject. More information regarding the segmentation of the DEMs by gray scale reconstruction usinggeodesic dilation can be found in [11] where similar algorithms are employed for extracting 3D objects

Remote Sens. 2011, 3 1331

as well as the ridge lines from high resolution LIDAR DSM. In a segmentation algorithm based ongeodesic reconstruction, selecting an appropriate “marker” image plays the main role and has a directeffect on the quality of the final reconstructed image. A “marker” image with a small offset, e.g., fewmeters, from the “mask” can suppress mainly local maxima regions similar to artifacts above the ground.

The proposed outlier extraction algorithm regarding the positive artifacts and anomalies is representedin Figure 7. The first step for segmentation based outlier detection is to generate the “marker” image as asecond input image. The first input image is the original ASTER GDEM as “mask” image. The markeris generated by subtraction of an offset value h from the ASTER GDEM:

mask+ = GDEM

marker = mask+ − h

where mask+ is the “mask” image employed to extract the “positive” outliers. Since the artifacts are inmost cases located far beyond the ground elevation level with low height variation of their internal pixels,a single offset value (h) of about “25 m” generates an appropriate marker image for segmentation of theoutliers for ASTER GDEM. This is a basic assumption for the outlier elimination algorithm. It means ifan outlier region contains low inclination on most parts of its boundary to the neighboring ground pixels,the algorithm might fail to detect it.

Figure 7. Proposed workflow for extracting positive outliers.

Figure 8 shows a small ASTER GDEM tile with 420 × 550 pixels which is used as “mask” inthis algorithm. After providing the second input image (“marker”), the image reconstruction ImRecis determined accordingly. For segmentation purpose, the reconstructed image is subtracted from theoriginal DEM. The result is a normalized DEM similar to normalized Digital Surface Model (nDSM)where the regions which are about h meter higher than their neighborhood are highlighted. Similarresults can be achieved by morphological “tophat” filtering [12,13], but the operation based on geodesic

Remote Sens. 2011, 3 1332

dilation is better suited because of its independence from the size of the objects to be filtered. Therefore,there is no need to tune the size of the structuring element.

Figure 8. ASTER GDEM sample data corresponding to an area of 24 km × 39 km.

40

50

60

70

80

90

100

The segmentation procedure is implemented using thresholding and the labeled regions are providedby means of connected components analysis. A geometric feature descriptor is created which highlightsthe height variation on each pixel regarding its adjacency to evaluate the labeled regions. A feature calledLocal Range Variation (LRV) is created by subtracting the maximum and minimum values in every 3×3windows over the image (cf. Figure 9). All the boundary pixels of the detected regions are evaluated bythe LRV descriptor. The regions having certain height jumps on their boundary pixels will be evaluatedas outliers (positive artifacts). In practice, the LRV values of the boundary of each region are extractedand if the majority (here 90%) of LRV values are above the threshold (here above 25 m), the region isclassified as outlier and corresponding pixels are eliminated from the original data set.

The pixel values corresponding to the LRV are additionally utilized to determine the values thatare taken by h (offset from DSM) for automatic generation of the marker image in each iteration.Accordingly the process begins by choosing the maximum value of LRV as initial offset h for markergeneration. For an efficient extraction of all outliers, ten offsets are determined by dividing the differencebetween maximum and minimum LRV into ten equal intervals. The iteration begins with maximum valueand the provided segments based on proposed algorithm are evaluated accordingly. Procedure continuesby selecting the next h and evaluating the new segments which are not evaluated in previous step untilno more new segments are produced. A similar process is utilized to eliminate the negative outliers butin this case, the complementary image of the ASTER GDEM is selected as “mask” and therefore:

mask− = max(GDEM)−GDEM

Remote Sens. 2011, 3 1333

marker = mask− − h

where mask− is used to extract the negative outliers and is created by inverting the original GDEM.Classified outlier regions in this step are then integrated and corresponding pixels are eliminated fromthe original ASTER GDEM. Figure 10 illustrates the final detected outliers from the example ASTERGDEM image.

Figure 9. Local Range Variation (LRV) feature descriptor.

0

5

10

15

20

25

30

Figure 10. Final detected positive and negative outliers.

The segment-based algorithm is employed to eliminate the area-shaped artifacts such as “pits” and“bumps”. An additional filtering is integrated to reduce the “mole-runs” errors. Since the “mole-runs”

Remote Sens. 2011, 3 1334

appear as curvilinear and positive height errors, a morphological opening filter is utilized to eliminatethem. A disk shaped structuring element with the radius of 7 pixels is selected to suppress curvilinearelevated errors (cf. Figure 11). Median filter is a well-known filtering method of this category, whichreplaces the value of a pixel by the median of gray values in the neighborhood (cf. Figure 11).

Figure 11. Removal of the curvilinear height errors of “mole-run”. (a) ASTER GDEM;(b) Median filter using a neighborhood image size of 7 × 7; (c) Morphology Opening usinga disc shape of radius length of 3 pixels.

(a) (b) (c)

Figure 11 represents the corresponding results by applying median filter and morphological openingon original ASTER GDEM. For the median filter, a neighborhood image size of 7× 7 is selected for thisexample which depends on the width of the linear errors to be suppressed. A similar size of structuringelement is used for morphological opening to be comparable with the result of median filter. In this work,a disc shaped structuring element with the radius of 3 is employed. To assess the quality of the employedfilters for eliminating the mole-runs, a profile plot is provided (cf. Figure 12). In this figure, the blackline corresponds to the original ASTER GDEM points where the locations of the several mole-runs aspositive peaks are visualized. Median filter (blue line) suppresses parts of the peaks as well as smoothingthe other terrain parts. Since the aim is to suppress the sharp peaks with low interference with the other

Remote Sens. 2011, 3 1335

pixels, image opening using an appropriate structuring size creates a better result (cf. Figure 12, redline).

Figure 12. Profiles from images represented in Figure 11, showing that the median filterdoes not suppress properly the mole-runs errors

2.3. Correction of the Height Errors of Water Body Regions:

A further criterion is considered for the final classification into outlier and inliers regions. Duringproduction of the ASTER GDEM, the pixels inside inland water bodies are not filtered and, therefore,they contain random errors. In this step, a water mask binary image is provided as a sort of qualitylayer that warns the user about the lower quality of height points inside water body areas. In addition,the water mask layer can directly be used to filter out all the height points inside the water. In order toprovide high quality water mask, the vector map containing the boundary points of the shoreline regionsin different resolutions are automatically extracted from the ”Global Self-consistent, Hierarchical,High-resolution Shoreline Database“ (GSHHS) which is is freely available to download [14]. GSHHS iscombined from two main databases World Data Bank II (WDB) and World Vector Shoreline (WVS). Theshorelines are provided as closed polygons and are free of intersections or other artifacts caused by datainaccuracies [15].

Figure 13(a) shows a small tile of ASTER GDEM corresponding to an area containing a water regionwith low elevation quality. Comparing to the SRTM image cf. Figure 13(b)), the different blue tonesin the water region illustrated that the water region is a merged result of smaller DEMs created fromdifferent numbers of stereo scenes (stack numbers). Particularly, such effects occur when the mergedDEMs have been generated using a small numbers of stereo scenes. In such situation the quality ofSRTM DEM (cf. Figure 13(b)) for the region with small stack numbers are much higher than the ASTERGDEM (cf. Figure 13(a)).

Accordingly, the height values inside the water regions in ASTER GDEM are flattened using thewater mask created from the GSHHS data set (cf. Figure 13(c)). After extracting the correspondingwater regions, the median value of the associated heights of the boundary region is inserted for the pixelsin the related water region. It should be mentioned that, the median of the heights is not applied for

Remote Sens. 2011, 3 1336

rivers. Hence, the difference between the maximum and minimum heights of the boundary pixels aremeasured. If the difference is less than a predefined value, e.g., 2 m, the median of the heights of theboundary pixels is used as replacement, otherwise the pixels in the region remain intact.

Figure 13. ASTER GDEM, SRTM, and water mask provided by GSHHS data(3 km × 1.4 km).

30

40

50

60

70

80

90

(a) ASTER GDEM

45

50

55

60

65

70

75

80

85

(b) SRTM DEM (c) water mask

An additional check is required to guarantee that the new water level is not higher than the surroundingground pixels. In case of a higher water level, the median of the lowest 10% of the boundary pixelsis considered as water elevation. In the final step, the gaps provided by all the different outliers arefilled by applying a spatial interpolation procedure. In this paper, Inverse Distance Weighting (IDW)interpolation [16] is utilized.

3. Results

3.1. Test Data

The proposed algorithms for enhancement of the ASTER GDEM were experimentally tested ona dataset belonging to an area near Barcelona, Spain. The area is selected due to the availabilityof an accurate ground truth for final evaluation of the corrected DEM. The area is a mountainousregion containing 2,378 × 3,601 pixels with a height range from zero up to 1202 m. The data at thesoutheast part of the scene represent the Mediterranean Sea, which is visible by a very dark gray value(cf. Figure 14).

Remote Sens. 2011, 3 1337

Figure 14. Original ASTER GDEM; image size = 2378 × 3601 pixels corresponding to71.3 km × 108 km; Height variation = [0 m, 1202 m].

3.2. ASTER GDEM Enhancement

In the first enhancement step as explained in Section 2.1, a correction layer is provided whichrepresents the vertical offset for each pixel of the GDEM. For this purpose, the ICESat laser altimetrypoints corresponding to the test area are extracted. A refinement strategy on ICESat waveform laser datais accomplished to reduce the outliers and the laser points located on the top of trees. An additionalrefinement is considered to eliminate the points with a too high discrepancy to the ASTER GDEM forwhich a maximum deviation of about 50 m is used. Next, the correction layer is provided by spatialinterpolation of the differences between the remaining laser points and the corresponding GDEM points.The correction layer is then employed to shift the original ASTER GDEM pixels according to the valuesof the correction layer. The second enhancement step is dedicated to the segment-based outlier detectionprocedure. In this step the common artifacts, which are mentioned in Section 2.2 are detected andeliminated. The curvilinear elevation error, i.e., “mole-run” is reduced by employing a morphologicalopening filter and the two other errors, “pits” and “bumps”, are eliminated using the segment-basedalgorithm (cf. Figure 15). As a final process, the gaps created due to outliers are filled by means ofspatial interpolation using Inverse Distance Weighting (IDW) [16].

As the third enhancement step, the inland water bodies are extracted from the GSHHS vector database,as explained in Section 2.3. A water mask layer is created by converting the water polygons into a rasterimage. The process continues by filling corresponding water pixels of GDEM by the median values ofthe bordering pixels if the region is not regarded as a river.

Remote Sens. 2011, 3 1338

Figure 15. Outlier regions corresponding to the area indicated in Figure 14 with red polygon;ASTER GDEM (top) and corresponding outliers (bottom).

3.3. Comparison to the Reference DEM

The primary reference dataset used in this evaluation is a bare earth Digital Terrain Model (DTM)produced by the Institut Cartographic de Catalunya (ICC). It has a grid spacing of 15 m and an absolutevertical error of 1.1 m (RMSE). In order to compare it with ASTER GDEM, it has been resampled into30 m pixel size using bilinear interpolation (cf. Figure 16).

As shown, the ICC ground truth image does not fully cover the whole test area and therefore only thecommon pixels are assessed in this section. As a first step, the quality of corresponding ICESat and ICCground truth points are evaluated against each other to check how they fit to each other. A histogramas well as statistical measures are provided to highlight the differences of these two reference datasets(cf. Figure 17 and Table 1). They prove that the ICC ground truth is of very high quality. The averagevariation of about 18 cm as well as the other parameters, particularly, quantile for 90% of the valuesdemonstrate an appropriate quality as well as a very low rate of outliers.

After evaluating the ground truth the quality of all corrected GDEM points is analyzed against the ICCground truth. The evaluation includes histograms and determines statistical parameters of the deviationsbetween input and reference datasets. In addition, a profile-view along a specific direction of the ASTERGDEM before and after correction together with the ICC ground truth data shows the improvementvisually (cf. Figure 18). In the profiles in Figure 18, the blue lines corresponds to the ICC ground truthdata, the red lines show the location of the original ASTER GDEM pixels, and the black lines representthe position of the corrected GDEM pixels.

Remote Sens. 2011, 3 1339

Figure 16. ICC Ground truth DEM.

Figure 17. Difference between the ICESat points and ground truth ICC data.

Obviously, the location of the GDEM points after correction is shifted towards the ICC points withmuch better correspondence. Also local improvements of the DEM shape can be seen in the profile ofthe corrected data.

Remote Sens. 2011, 3 1340

Table 1. Comparison of the ground truth against ICESat points.

Diff. = ICC – ICESat

Min –4.66 m

Max 4.89 m

Mean 0.18 m

Median 0.13 m

Standard Deviation 0.85 m

RMSE 0.87 m

Quantile 90% 1.01 m

Figure 18. Profiles compare the location of the ASTER GDEM pixels before (red) and after(black) correction as well as the ICC ground truth (blue).

Figure 19 displays the histograms of the differences between the corrected ASTER GDEM from theICC points as blue bars overlaid on the histogram of the differences between the original ASTER GDEMand the ICC pixels as red bars. The histograms are corresponding to all the pixels in the common area ofthe images. The comparison of the two histograms confirms the overall movement of the bias towards avery low value close to zero.

Table 2 represents the statistical measures evaluating the quality of the corrected ASTER GDEMin comparison to the original GDEM. It is obvious that the mean and median measures show almostthe same improvement as for the comparison against the ICESat points. The table also shows that themaximum and minimum differences have significantly improved, which is due to eliminating the worst

Remote Sens. 2011, 3 1341

outliers from the scene. Although the standard deviation is reduced by only 0.23 m, this still means thatseveral regions with outliers must have been reduced, which can be partly seen in Figure 18.

Figure 19. Histograms displaying the differences between original ASTER GDEM and ICCground truth (red) as well as the differences between corrected ASTER GDEM and ICCground truth.

Table 2. Statistical parameters representing the quality improvement of the correctedASTER GDEM. The ASTER GDEMs before and after correction is compared against theICC ground truth.

OrigGDEM – ICC CorrGDEM – ICC

Min –193.85 m –57.36 m

Max 140.05 m 83.66 m

Mean –12.99 m 0.02 m

Median –12.78 m 0.03 m

Standard Deviation 8.21 m 7.98 m

RMSE 15.36 m 7.98 m

Quantile 90% –3.87 m 9.16 m

Remote Sens. 2011, 3 1342

4. Conclusions

In this paper, an algorithm consisting of three main steps is proposed for enhancing the quality of theASTER GDEM. In the first step, the ICESat laser altimetry data are used to provide a correction layer toreduce the height offset of the global DEM.

The second improvement step consists of the reduction of artifacts and outliers from the GDEM,which is performed based on a segment-based algorithm and a smoothing filter. The last step is to flattenthe inland water levels by generating a water mask based on the GSHHS shoreline vector dataset. Theevaluation based on ground truth data shows that the overall accuracy of the final corrected GDEM issignificantly improved and most of the errors and artifacts are removed. Additionally, the water elevationlevels are corrected. Moreover, it can be concluded that the ICESat laser points are an appropriatereference data for height correction. In further investigations it is envisaged to correct larger parts ofthe ASTER GDEM, especially above 60◦ North and below 58◦ South and combine SRTM and ASTERGDEM for remaining areas.

References

1. ERSDACT. Earth Remote Sensing Data Analysis Center. Available online: http://www.ersdac.or.jp(accessed on 7 June 2011).

2. ERSDACT(a). ASTER Global DEM Validation Summary Report; Technical Report; Availableonline: http://www.gdem.aster.ersdac.or.jp (accessed on 7 June 2011).

3. Huber, M.; Wessel, B.; Kosmann, D.; Felbier, A.; Schwieger, V.; Habermeyer, M.; Wendleder, A.;Roth, A. Ensuring Globally the TanDEM-X Height Accuracy: Analysis of the Reference Data SetsICESat, SRTM and KGPS-Tracks. In Proceedings of IEEE International Geoscience and RemoteSensing Symposium (IGARSS ’09), Cape Town, South Africa, 12–17 June 2009; pp. 769-772.

4. Duong, H.; Lindenbergh, R.; Pfeifer, N.; Vosselman, G. ICESat full-waveform altimetry comparedto airborne laser scanning altimetry over The Netherlands. IEEE Trans. Geosci. Remote Sens.2009, 47, 3365-3378.

5. NASA/ICESat. Available online: http://icesat.gsfc.nasa.gov/ (accessed on 7 June 2011).6. Rodriguez, E.; Morris, C.; Belz, J.; Chapin, E.; Martin, J.; Daffer, W.; Hensley, S. An Assessment

of the SRTM Topographic Products; Technical Report JPL D-31639; Jet Propulsion Laboratory:Pasadena, CA, USA, 2005; p. 143.

7. Arefi, H.; Reinartz, P. Elimination of the Outliers from Aster GDEM data. In Proceedings ofThe 2010 Canadian Geomatics Conference and Symposium of Commission I, ISPRS, Calgary, AB,Canada, 15–18 June 2010; In IAPRS; 2010; Vol. 38, Part 1.

8. Abrams, M.; Bryan, B.; Hiroji, T.; Masami, H. The ASTER Global DEM. Photogramm. Eng.Remote Sensing 2010, 76, 344-348.

9. Vincent, L. Morphological grayscale reconstruction in image analysis: Applications and efficientalgorithms. IEEE Trans. Image Process. 1993, 2, 176-201.

Remote Sens. 2011, 3 1343

10. Arefi, H.; Hahn, M. A Morphological Reconstruction Algorithm for Separating Off-Terrain Pointsfrom Terrain Points in Laser Scanning Data. In Proceedings of the ISPRS Workshop LaserScanning 2005, Enschede, The Netherlands, 12–14 September 2005; In IAPRS; 2005; Vol. 36,Part 3/W19.

11. Arefi, A. From LIDAR Point Clouds to 3D Building Models. Ph.D. Thesis, BundeswehrUniversity, Munich, Germany, 2009.

12. Gonzalez, R.C.; Woods, R.E. Digital Image Processing; Prentice Hall: Upper Saddle River, NJ,USA, 2008.

13. Soille, P. Morphological Image Analysis: Principles and Applications, 2nd ed.; Springer-Verlag:Secaucus, NJ, USA, 2003.

14. GSHHS—A Global Self-consistent, Hierarchical, Highresolution Shoreline Database; Availableonline: http://www.ngdc.noaa.gov/mgg/shorelines/gshhs.html (accessed on 7 June 2011).

15. Wessel, P.; Smith, W.H.F. A global self-consistent, hierarchical, high-resolution shoreline database.J. Geophys. Res. 1996, 101, 8741-8743.

16. Haining, R. Spatial Data Analysis: Theory and Practice, 1st ed.; Cambridge University Press:Cambridge, UK, 2003.

c© 2011 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access articledistributed under the terms and conditions of the Creative Commons Attribution license(http://creativecommons.org/licenses/by/3.0/.)